the deeper point here connects to something the field keeps rediscovering. we trained reasoning models to think longer. then we discovered longer doesn't mean better. now this paper shows the models themselves already know that. they're generating stop signals that our inference
GENERATIVE AI
-
DeepSeek and Qwen3 performance improvements
By
–
specific numbers worth sitting with: > DeepSeek-R1-7B on MATH-500: 93% accuracy (up from 91.6%), tokens cut from 3,871 to 2,141 > DeepSeek-R1-1.5B on AIME 2025: accuracy jumps 6.2 percentage points > Qwen3-8B: response length halved from 18,342 to 9,183 tokens with no accuracy
-
SAGE-RL: Simple RL Modification for Better Training
By
–
then they fold it into training with SAGE-RL. dead simple modification: in standard reinforcement learning (GRPO), you sample 8 responses per question. SAGE-RL replaces 2 of those 8 with SAGE-generated samples. the other 6 stay normal. one-line code change. the model learns to
-
SAGE: Efficient Step-by-Step Reasoning
By
–
their solution: SAGE (Self-Aware Guided Efficient Reasoning). instead of generating token by token, SAGE extends chains in whole reasoning steps. after each step, it checks: is the model confidently signaling it wants to stop? if yes, reasoning ends. no fine-tuning. no new
-
Three Key Observations on Model Paths
By
–
three observations back this up: > high-confidence paths (measured by cumulative probability) are shorter AND more accurate than standard samples > at those natural endpoints, the stop token is the model's top prediction. it knows it's finished > as you widen exploration (more
-
Researchers test AI self-awareness with TSearch
By
–
here's where it gets interesting. the researchers probed whether models internally "know" they're done. they introduced TSearch, which scores partial reasoning traces by cumulative log-probability across the entire chain, not just the next token. when you let the model explore
-
Overthinking hurts accuracy in AI responses
By
–
and it's not just wasted compute. overthinking actively hurts accuracy. DeepSeek-R1 produces responses 5x longer than Claude 3.7 Sonnet on AIME 2025 with comparable accuracy. QwQ-32B scores 2 percentage points HIGHER with its shortest answers using 31% fewer tokens. 72% of
-
RFCS Metric Reveals Early Correct Steps in Models
By
–
first, the problem quantified. the researchers created a metric called RFCS (Ratio of First Correct Step) that tracks where in a chain of thought the correct answer first appears. on MATH-500, across every model tested, the right answer shows up well before the end in over half
-
Overthinking in AI models is a sampling issue
By
–
reasoning models already know when they've solved the problem. we just don't let them stop. new paper from Beihang University and ByteDance shows that the overthinking problem in models like DeepSeek-R1 and Qwen3 isn't a training failure. it's a sampling failure. the fix cuts
-
Anthropic Revenue Growth, Contract Compromise and Nvidia Results
By
–
3-4-2026 ($) •Anthropic's Skyrocketing Revenue
•A Contract Compromise?
•Nvidia Earnings https://stratechery.com/2026/anthropics-skyrocketing-revenue-a-contract-compromise-nvidia-earnings/ [Translated from EN to English]→ View original post on X — @benthompson, 2026-03-04 11:00 UTC